Covid-19 Detection from Chest X-Ray Images using a Low-Tier Android Mobile Phone with Lightweight Transfer Learning Networks

Date of Award

12-2021

Document Type

Thesis

Degree Name

Master of Science in Data Science

First Advisor

Patricia Angela R. Abu, PhD

Abstract

The efforts to inoculate majority of the population have been slower than expected and this is especially true for low income countries. This problem further accentuates the importance of effective mass testing, con- sidering the emergence of newer variants. The RT-PCR is still the gold standard diagnostic test for COVID-19 detection, but its limitations have led researchers to explore supplementary screening methods. Chest X-Ray imaging is one potential tool to consider, and its combination with deep learning for automation has piqued attention from the artificial intelligence community. As a further contribution to the field, this work focuses on creat- ing, evaluating, and comparing mobile-phone-suitable COVID-19-detecting models. These networks together with their corresponding quantized ver- sions were tested for their classification performance, resource consump- tion, and latency measurements when pushed to a low-tier android device. Results show that the utilization of EfficientNetB0, MobileNetV3Large, Mo- bileNetV3Small, and NASNetMobile architectures for the transfer learning task without any quantization are expected to generate 92.107%, 91.150%, 90.677%, and 86.314% overall average accuracy respectively for the 3-class classification scheme. For systems requiring more efficient models, utiliz- ing the quantized versions of the transfer learning models particularly with EfficientNetB0 or MobileNetV3Large as foundation is advantageous, as it renders only 0.116% and 0.009% accuracy loss respectively while maintain- ing more than 95% F1-scores for the COVID-19 class.

Link to Full Text

Share

COinS